Source-linked AI summary
Low-Power, Neuromorphic, Acoustic Anomaly Detection for Persistent Machine Monitoring
Steven C. Nesbit, Victor M. Vergara, Michael A. Felix, Evan T. Kain, Luis R. García Carrillo, Gerd J. Kunde, Andrew T. Sornborger
TL;DR
Persistent acoustic machine monitoring needs always-on fault detection without contact, but power and deployment constraints limit inference. This paper implements log-mel autoencoder anomaly detection on Loihi 2, achieving strong clean-benchmark separation and exceeding the DCASE ToyCar baseline in noisy conditions.
Problem
Persistent acoustic monitoring offers noncontact fault detection, but always-on inference remains constrained by power, latency, and deployment complexity.
Method
The study performs one-class log-mel autoencoder anomaly detection with reconstruction scoring and thresholding implemented on a Loihi 2 neuromorphic processor.
Results
The system achieves 0.9959 AUC on clean ToyCar and exceeds the DCASE 2026 ToyCar baseline across all three reported metrics in noisy conditions.
Takeaways & Limitations
These results support Loihi 2 as a practical substrate for low-power, persistent acoustic anomaly detection across clean and noisy monitoring conditions.
Takeaways & Limitations
Log-mel extraction remains off chip, and direct single-chip Loihi 2 measurements and broader field evaluations remain future work.
Abstract
from arXiv · showhide
Persistent acoustic monitoring can detect machine faults without physical contact, but always-on inference is constrained by power, latency, and deployment complexity. We demonstrate autoencoder-based acoustic anomaly detection on an Intel Loihi 2 neuromorphic processor under clean and noisy conditions. Log-mel features are computed off chip; normalization, autoencoder inference, L1 reconstruction scoring, and thresholding run on chip. In a clean, microphone-position-invariant ToyADMOS ToyCar benchmark, the on-chip model achieves 0.9959 AUC and 0.9785 standardized pAUC at maximum false-positive rate 0.1. In the DCASE 2026 Task 2 ToyCar noisy benchmark, the model achieves source AUC 0.7990, target AUC 0.6466, and pAUC 0.6426, exceeding reported baseline metrics. Power profiling on a 16-chip Loihi 2 VPX system shows real-time throughput with 0.0406$\unicode{x2013}$0.0426 mJ dynamic energy per sample, two orders of magnitude lower than both a CPU and GPU. These results support neuromorphic acoustic anomaly detection as a practical candidate for low-power, persistent machine monitoring.
1. INTRODUCTION
Persistent acoustic monitoring offers noncontact detection of rare machine faults, but requires energy-efficient always-on inference. This study implements strict one-class autoencoder anomaly detection on Loihi 2, with log-mel extraction off chip and normalization, inference, scoring, and thresholding on chip, evaluated on clean and noisy benchmarks with hardware profiling.
- Motivation: Acoustic monitoring enables noncontact observation of rotating and vibrating equipment because faults can alter measurable spectral content.Persistent monitoring must remain active despite mostly nominal operation so rare faults can be identified quickly.
- Motivation: Machine-condition monitoring suits one-class detection because failures are rare, heterogeneous, and expensive to collect, motivating normal-sound benchmarks such as ToyADMOS, MIMII, and DCASE.These benchmarks train models primarily or exclusively on normal machine sounds.
- Approach: The study trains both autoencoders exclusively on normal machine sounds and uses anomalous recordings only for evaluation.This strict one-class formulation targets previously unseen failures rather than relying on known fault labels.
- Approach: Log-mel extraction runs off chip, while normalization, dense autoencoder inference, L1 reconstruction scoring, and thresholding run on Loihi 2.The persistent-monitoring pipeline assigns feature extraction off chip and anomaly decisions on chip.
- Evaluation: The evaluation covers clean ToyADMOS and noisy DCASE 2026 ToyCar data, together with latency, power, energy, memory, and activity profiling against server-grade CPU and GPU implementations.The contribution targets neuromorphic on-chip autoencoder inference and anomaly decisions across clean and noisy conditions.
2. METHODS
The study evaluates neuromorphic autoencoder anomaly detection on clean and noisy ToyCar recordings using log-mel features and on-chip fixed-point reconstruction scoring. Clean data pools four microphone positions, while noisy evaluation uses only near-channel recordings from the corrected DCASE 2026 release [17].
- Datasets: Clean evaluation uses ToyADMOS ToyCar recordings pooled across four microphone positions, with 21 600 normal and 4 236 anomalous recordings.The training set contains 17 280 normal-condition recordings.
- Datasets: Noisy evaluation uses 16-kHz, 10- to 12-second DCASE 2026 ToyCar development recordings [17], retaining only near-channel audio.Each recording includes synchronized near and far channels, but the far channel is used only as a potential noise reference.
- Model and deployment: Both conditions train autoencoders on normal features, normalize with training-set Z-scores, convert inference to fixed-point arithmetic, and use reconstruction error for anomaly scoring.The clean and noisy bottlenecks contain 32 and 12 neurons, respectively, with leaky-ReLU hidden layers and linear outputs.
- Evaluation: Anomaly scores are computed from central one-second segments, with larger reconstruction scores indicating greater deviation; performance uses AUC and standardized pAUC at maximum FPR 0.1.The final score or a thresholded binary decision can be emitted directly by Loihi 2.
3. RESULTS
The quantized Loihi 2 autoencoder achieves strong clean-benchmark separation and exceeds the DCASE 2026 ToyCar baseline on the reported metrics. It also delivers real-time inference and substantially lower dynamic energy than CPU and GPU references, with measured resource usage and first-order power projections reported for deployment context.
- Detection results: 0.9959 AUC and 0.9785 standardized pAUC demonstrate strong clean-benchmark separation using the deployed L1 reconstruction score.The clean benchmark evaluates the quantized Loihi 2 autoencoder on normal and anomalous log-mel features.
- Detection results: 0.6466 target AUC exceeds the DCASE 2026 ToyCar baseline of 0.5317, while the Loihi 2 model exceeds the baseline on all three reported metrics.The strongest reported gain is on target AUC, improving from 0.5317 to 0.6466.
- Efficiency and deployment: 281.0 µs per sample with Ethernet I/O supports real-time monitoring, while 48.3 µs preloaded execution is faster than both CPU and GPU measurements.The measurements exclude feature extraction; GPU host-device transfer was also excluded.
- Efficiency and deployment: 0.0406–0.0426 mJ dynamic energy per sample is 474–496× lower than the CPU and 126–132× lower than the GPU, depending on communication handling.Loihi 2 uses 0.0426 mJ with Ethernet I/O and 0.0406 mJ with preloaded inputs, versus 20.16 mJ on the CPU and 5.35 mJ on the GPU.
- Efficiency and deployment: The mapped autoencoder occupies 74 neuromorphic cores, with 20.36% single-chip memory utilization and 1.27% utilization across the 16-chip VPX system.Per inference, probes recorded approximately 3.0 × 10^5 synaptic operations, 1.0 × 10^4 neuron updates, and 2.0 × 10^4 input and output spike events.
- Efficiency and deployment: Projected total power is 1.93 W with Ethernet I/O and 2.62 W with preloaded inputs, corresponding to 0.54 mJ and 0.13 mJ per sample.These are first-order projections based on estimated single-chip static power and assumed comparable dynamic and communication behavior on the smaller platform.
4. DISCUSSION
The results support Loihi 2 as a practical substrate for persistent acoustic anomaly detection, with useful anomaly-score ranking under clean microphone variation and noisy monitoring conditions. Low dynamic energy and limited core usage leave capacity for expansion, while off-chip log-mel extraction remains the principal limitation.
- Overall findings: Loihi 2 provides useful anomaly-score ranking across clean microphone-position variation and noisy monitoring using only the near microphone, without a separate noise-reference channel.The discussion characterizes the platform as practical for persistent acoustic anomaly detection across both monitoring conditions.
- Resource use: The model occupies 74 of approximately 2048 available neuromorphic cores, leaving capacity for additional models, sensor fusion, or future log-mel preprocessing.Static platform power dominates the 16-chip VPX measurement, but dynamic energy per sample is low.
- Limitations and future work: Log-mel extraction remains off chip, limiting host independence and total system-energy reduction despite Ethernet-I/O timing compatible with real-time monitoring.On-chip preprocessing could further reduce host dependence and total system energy.
- Limitations and future work: Future work should evaluate additional DCASE machine classes and field recordings, directly profile a single-chip Loihi 2 platform, and investigate online adaptation under changing machine and environmental conditions.These evaluations address broader deployment conditions and adaptation beyond the reported monitoring setup.
5. CONCLUSION
The paper demonstrates Loihi 2 neuromorphic log-mel autoencoder anomaly detection for clean and noisy acoustic monitoring. The system combines strong benchmark performance with real-time throughput and very low dynamic energy per sample compared with CPU and GPU inference.
- 5. CONCLUSION: 0.9959 AUC and 0.9785 pAUC are achieved in the clean acoustic anomaly-detection benchmark.The result comes from the Loihi 2 neuromorphic implementation.
- 5. CONCLUSION: Source AUC 0.7990, target AUC 0.6466, and pAUC 0.6426 exceed the DCASE 2026 ToyCar baseline.These results are reported for noisy acoustic monitoring.
- 5. CONCLUSION: Power profiling shows real-time throughput and very low dynamic energy per sample compared with CPU and GPU inference.This supports the implementation’s suitability for persistent acoustic monitoring.
7. COMPLIANCE WITH ETHICAL STANDARDS
The study used publicly available machine-sound datasets, so ethical approval was not required.
- 7. COMPLIANCE WITH ETHICAL STANDARDS: Ethical approval was not required because the study used publicly available machine-sound datasets.